Conventional disk drives have employed peak detection techniques in order to recover digital data written as saturation recording onto a magnetizable surface media of a rotating disk. With peak detection techniques, it is necessary to space flux transitions sufficiently apart so that analog peaks in the recovered data stream may be identified and the corresponding data recovered. In order to achieve reasonable bandwidths in data channels, it has been customary to employ data coding techniques. One such technique has been to use a (1,7) RLL code. In this code, flux transitions can be no closer together than every other clock bit time period ("bit cell") nor farther apart than eight clock bit cells. (1,7) RLL codes are known as "rate two-thirds" codes, in the sense that two data bits are coded into three code bits. Thus, with a rate two-thirds code, one third of the user storage area of the storage disk is required for code overhead.
One way to decrease the code overhead is to employ a code in which flux transitions are permitted in adjacent bit cells. One such code is a (0,4,4) code. The (0,4,4) code is generally thought of as a rate eight-ninths code, meaning that nine code bits are required for eight incoming data bits. (Theoretically, the (0,4,4) code ratio is somewhat higher, approaching 0.961) Thus, this code is significantly more efficient than a rate two-thirds code, such as (1,7) RLL. Use of a (0,4,4) code results in a significantly greater net user data storage capacity on the disk surface, given a constant bit cell rate. However, when flux transitions occur in adjacent bit cells, as is the case with a (0,4,4) code, intersymbol interference ("ISI") results. Conventional peak detection techniques are not effective or reliable in recovering data coded in an eight-ninths code format, such as (0,4,4).
The zero in the (0,4,4) code denotes that flux transitions may occur in directly adjacent bit cells of the coded serial data stream. The first "4" denotes that a span of no more than four zeros occurs between ones in the encoder output. The second "4" signifies that the bit cell stream has been-divided into two interleaves: an even interleave, and an odd interleave; and, it denotes that there can be a span of no more than four zeros between ones in the encoder output of either the odd interleave or the even interleave.
It is known that partial response signalling enables improved handling of ISI and allows more efficient use of the bandwidth of a given channel. Since the nature of ISI is known in these systems, it may be taken into account in the decoding/detection process. Partial response transmission of data lends itself to synchronous sampling and provides an elegant compromise between error probability and the available spectrum. The partial response systems described by the polynomials 1+D, 1-D, and 1-D.sup.2 are known as duobinary, dicode and class IV (or "PR4"), respectively, where D represents one bit cell delay and D.sup.2 represents 2 bit cell delays (and further where D=e.sup.-j.omega.T, where .omega. is a frequency variable in radians per second and T is the sampling time interval in seconds). The PR4 magnitude response plotted in FIG. 1 hereof and given the notation .vertline.1-D.sup.2 .vertline. emphasizes midband frequencies and results in a read channel with increased immunity to noise and distortion at both low and high frequencies. In magnetic recording PR4 is a presently preferred partial response system, since there is a close correlation between the idealized PR4 spectrum as graphed in FIG. 1, and the natural characteristics of a magnetic data write/read channel.
In order to detect user data from a stream of coded data, not only must the channel be shaped to a desired partial response characteristic, such as the PR4 characteristic, but also a maximum likelihood ("ML") sequence estimation technique is needed. The maximum likelihood sequence estimation technique determines the data based upon an analysis of a number of consecutive data samples taken from the coded serial data stream, and not just one peak point as was the case with the prior peak detection methods.
One maximum likelihood sequence estimation algorithm is known as the Viterbi detection algorithm, and it is well described in the technical literature. Application of the Viterbi algorithm to PR4 data streams within a magnetic recording channel is known to improve detection of original symbol sequences in the presence of ISI and also to improve signal to noise ratio over comparable peak detection techniques.
In an article entitled "Viterbi Detection of Class IV Partial Response on a Magnetic Recording Channel" appearing in IEEE Trans. on Communications. vol. Com-34, No. 5, May, 1986, pp. 434-461, authors Wood and Peterson explain the derivation of PR4 as being formed by subtracting waveforms two bit intervals apart, thereby forming an analog domain ternary "eye" pattern graphed herein in FIG. 2.
The Viterbi algorithm provides an iterative method of determining the "best" route along the branches of a trellis diagram, such as the one shown in FIG. 3 hereof, for example. If, for each trellis branch, a metric is calculated which corresponds to the logarithm of the probability for that branch, then the Viterbi algorithm may be employed to determine the path along the trellis which accumulates the highest log probability, i.e., the "maximum likelihood" sequence. Since the Viterbi algorithm operates upon a sequence of discrete samples {Yk}, the read signal is necessarily filtered, sampled, and equalized.
While PRML has been employed in communications signalling for many years, it has only recently been applied commercially within magnetic hard disk drives. One recent application is described in a paper by Schmerbeck, Richetta, and Smith, entitled "A 27 MHz Mixed Analog/Digital Magnetic Recording Channel DSP Using Partial Response Signalling with Maximum Likelihood Detection", Proc. 1991 IEEE International Solid State Circuits Conference, pp. 136-137, 304, and pp. 96, 97 and 265 Slide Supplement. While the design reported by Schmerbeck et al. appears to have worked satisfactorily, it has drawbacks and limitations which are overcome by the present invention. One drawback of the reported approach was its design for transducers of the ferrite MiG type or of the magnetoresistive type which simplified channel equalization requirements. Another drawback was the use of a single data transfer rate which significantly simplified channel architecture. A further drawback was the use of a dedicated servo surface for head positioning within the disk drive, thereby freeing the PR4, ML data channel from any need for handling of embedded servo information or for rapid resynchronization to the coded data stream following each embedded servo sector.
Prior Viterbi detector architectures and approaches applicable to processing of data sample sequences taken from a communications channel or from a recording device are also described in the Dolivo et al. U.S. Pat. No. 4,644,564. U.S. Pat. No. 4,504,872 to Peterson describes a digital maximum likelihood detector for class IV partial response signalling. An article by Roger W. Wood and David A. Peterson, entitled: "Viterbi Detection of Class IV Partial Response on a Magnetic Recording Channel" IEEE Trans, on Comm. Vol. Com-34, No. 5, May, 1986, pp. 454-466 describes application of Viterbi detection techniques to a class IV partial response in a magnetic recording channel. An article by Roger Wood, Steve Ahigrim, Kurt Hallarnasek and Roger Stenerson entitled: "An Experimental Eight-Inch Disc Drive with One-Hundred Megabytes per Surface", IEEE Trans. on Magnetics, Vol. Mag-20, No. 5, September, 1984, pp 698-702 describes application of class IV partial response encoding and Viterbi detection techniques as applied within an experimental disk drive. A digital Viterbi detector capable of withstanding lower signal to noise ratios is described in Matsushita et al. U.S. Pat. No. 4,847,871. These documents are representative examples of the known state of the prior art.
It is known that equalization is an important part of any PR4,ML data channel. Without proper shaping, the digital samples entering the Viterbi detector will not occur at three distinct levels and produce the desired open eye pattern, FIG. 2. Ideally, a PR4,ML channel equalizer will provide for band limiting beyond the Nyquist frequency, will provide for magnitude adjustment and will provide for phase compensation. While recognized as needed, the equalization task is made more difficult when zoned data recording techniques, and embedded servo techniques are employed within a disk drive PR4,ML data channel.
One proposed solution is discussed for example in an article by J. D. Coker, R. L. Galbraith, G. J. Kerwin, J. W. Rae and P. A. Ziperovich entitled "Implementation of PRML in a Rigid Disk Drive" IEEE Trans. on Magnetics, Vol. 27, No. 6, November, 1991, pp. 4538-4543. The data channel described in that article included a three-tap FIR feed forward filter using a single K adjustment parameter. The center tap weight multiplier in the filter was maintained at unity while the two remaining tap weight multipliers were specified by a common K value. The transfer function of the filter was defined by: H(f)=1+2Kcos(2.pi.fT). This transfer function provided only magnitude adjustment and did not modify the phase of the signal except for fixed group delay errors. The Coker et al. approach was intended to compensate for differences in characteristics from the inside disk diameter to the outside diameter. A five-byte training sequence, written as part of each sector header format, was used to train the three tap FIR filter to an optimized response for the user data of the particular data sector. A logic algorithm function operating at the output of the FIR filter was used to generate a .increment.K error term. The .increment.K error term was digitally integrated to derive the K value which was used to adjust the transfer function of the FIR filter.
While the foregoing approach provided for some adaptivity and training for an FIR filter, it did so in an operating environment not including reconfiguration for embedded sector servo information written at a different data transfer rate than the user data, and also not including zone-data-recording techniques. Thus, a hitherto unsolved need has arisen for an improved FIR filter for adaptive equalization of a PR4,ML data channel, and for an improved filter training method for the FIR filter.